MoE load-balancing loss mathematical notation

MoE load-balancing loss is a recurring research-paper notation family. Penalizes router behavior that concentrates tokens on too few experts by coupling routing probability mass with actual assignment frequency.

MoE load-balancing loss: Penalizes router behavior that concentrates tokens on too few experts by coupling routing probability mass with actual assignment frequency. Example: Discourage experts that receive both high probability and disproportionate token load.